Uncertainty quantification for plant disease detection using Bayesian deep learning

Abstract Climate change is having an enormous impact on crop production in Latin America and the Caribbean. This problem not only concerns the volume of crop production but also the quality and safety of the food industry. Several research studies have proposed deep learning for plant disease detection. However, there is little information about the confidence of the prediction on unseen samples. Therefore, uncertainty in models of plant disease detection is required for effective crop management. In particular, uncertainty arising from sample selection bias makes it difficult to scale automatic plant disease detection systems to production. In this paper, we develop a probabilistic programming approach for plant disease detection using state-of-the-art Bayesian deep learning techniques and the uncertainty as a miss-classification measurement. The results show that Bayesian inference achieves classification performance that is comparable to the standard optimization procedures for fine-tuning deep learning models. At the same time, the proposed method approximates the posterior density for the plant disease detection problem and quantify the uncertainty of the predictions for out-of-sample instances.